Run the TrackQuality ANN directly from the onnx file with ONNX Runtime

[AndrewEdmonds11]

This PR contains a draft of the code needed to run the TrkQual ANN directly from the onnx file. This will remove TMVA::SOFIE from the workflow. This is a draft while I work out final details and remove the old code. However, it is validated:

Begin processing the 1st record. run: 1430 subRun: 0 event: 5 at 30-Mar-2026 10:45:40 CDT
[TrackQuality::produce::TrkQualAll] Inputs = 25, 1.0000, 1.2688, 0.2400, 0.6373, 0.1149 1.0800 --> output = 0.9459 (ORT: 0.9459)
[TrackQuality::produce::TrkQualAll] Inputs = 20, 0.8000, 1.4036, 0.4500, 0.0159, 0.1751 1.4000 --> output = 0.1627 (ORT: 0.1627)
[TrackQuality::produce::TrkQualAll] Inputs = 25, 1.0000, 1.2698, 0.4400, 0.1736, 0.2697 1.0800 --> output = 0.1440 (ORT: 0.1440)
[TrackQuality::produce::TrkQualAll] Inputs = 22, 0.8800, 1.3619, 0.4091, 0.0000, 0.1486 1.2273 --> output = 0.2698 (ORT: 0.2698)
Begin processing the 2nd record. run: 1430 subRun: 0 event: 8 at 30-Mar-2026 10:45:40 CDT
[TrackQuality::produce::TrkQualAll] Inputs = 42, 1.0000, 0.4454, 0.0952, 0.0000, 0.1808 1.0476 --> output = 0.8104 (ORT: 0.8104)
[TrackQuality::produce::TrkQualAll] Inputs = 32, 0.7619, 1.1694, 0.2500, 0.0000, 0.2672 1.3125 --> output = 0.1182 (ORT: 0.1182)
[TrackQuality::produce::TrkQualAll] Inputs = 42, 1.0000, 0.4456, 0.1429, 0.0000, 0.2196 1.1190 --> output = 0.5403 (ORT: 0.5403)
[TrackQuality::produce::TrkQualAll] Inputs = 34, 0.8095, 1.1090, 0.2353, 0.0000, 0.3372 1.2647 --> output = 0.0798 (ORT: 0.0798)
Begin processing the 3rd record. run: 1430 subRun: 0 event: 9 at 30-Mar-2026 10:45:40 CDT
[TrackQuality::produce::TrkQualAll] Inputs = 53, 0.7162, 0.8714, 0.3019, 0.0000, 0.0619 1.6226 --> output = 0.8246 (ORT: 0.8246)
[TrackQuality::produce::TrkQualAll] Inputs = 73, 0.9865, 0.4028, 0.0959, 0.0008, 0.0537 1.1233 --> output = 0.9948 (ORT: 0.9948)
[TrackQuality::produce::TrkQualAll] Inputs = 55, 0.7432, 0.8394, 0.3091, 0.0000, 0.0855 1.4727 --> output = 0.7908 (ORT: 0.7908)
[TrackQuality::produce::TrkQualAll] Inputs = 73, 0.9865, 0.4080, 0.0959, 0.2266, 0.0709 1.1233 --> output = 0.9954 (ORT: 0.9954)
[TrackQuality::produce::TrkQualAll] Inputs = 88, 0.9670, 0.4044, 0.0455, 0.0000, 0.0744 1.0227 --> output = 0.9959 (ORT: 0.9959)
[TrackQuality::produce::TrkQualAll] Inputs = 74, 0.8043, 0.7272, 0.2838, 0.0000, 0.0930 1.1757 --> output = 0.9649 (ORT: 0.9649)
[TrackQuality::produce::TrkQualAll] Inputs = 84, 0.9231, 0.7048, 0.1071, 0.0000, 0.1344 1.0952 --> output = 0.9896 (ORT: 0.9896)
[TrackQuality::produce::TrkQualAll] Inputs = 65, 0.7065, 0.8516, 0.2923, 0.0000, 0.1807 1.2923 --> output = 0.5042 (ORT: 0.5042)

where output is the original output and ORT is value with the new ONNXRuntime code

[oksuzian]

Took a look through this draft. Validation numbers match SOFIE to 4 decimals — behavior looks correct. A few structural items for when you come back to clean it up:

Blockers

• Hardcoded ONNX path.

  _session(_env, "ArtAnalysis/TrkDiag/data/TrkQual_ANN1_v2.onnx", _session_options),

  Only works when cwd is set correctly. The existing SOFIE path uses ConfigFileLookupPolicy(conf().datFilename()) — add a fhicl::Atom<std::string> onnxFilename and resolve it the same way. Since _session is initialized in the ctor init list, you'll need to resolve the path inline in the init-list expression (or via a small helper).

Bugs / risks

• print_shape is unused, and v.size() - 1 underflows to SIZE_MAX if v is empty — either wire it up behind _debug or delete.
• Member-init ordering is load-bearing and fragile. _env → _session_options → _session → _input_name → _type_info → _tensor_info → _input_shape → _memory_info → _output_name must appear in this order both in the class body and the ctor init list. Worth a one-line comment above the member declarations noting this.
• (double) casts before assignment into std::vector<float> are meaningless (value gets narrowed to float anyway). Either make the vector double or drop the casts.
• output_data[0] = 0 mutates the ORT tensor after Run. Works but surprising — float ort_score = entrance_found ? output_data[0] : 0.f; reads more naturally.

Nits

• SConscript indentation is inconsistent: surrounding lines use 2 spaces, new 'onnxruntime' uses 4 (mainlib) and 8 (plugins). Match surrounding style.
• _session_options and _allocator are only needed at construction. Could be locals. Minor.
• Dynamic-dim rewrite (if (dim == -1) dim = 1;) fixes batch=1. Batching N tracks per event would be a cheap win later — worth a TODO.

[AndrewEdmonds11]

This PR is now ready for merging. Responses to the Copilot review:

Blockers:

Hardcoded ONNX path.

Done in 95fe557

Bugs / risks:

print_shape is unused, and v.size() - 1 underflows to SIZE_MAX if v is empty — either wire it up behind _debug or delete.

Function removed in 0b918c0

Member-init ordering is load-bearing and fragile. ... Worth a one-line comment ...

Comment added in 8028fb4

(double) casts before assignment into std::vector are meaningless (value gets narrowed to float anyway). Either make the vector double or drop the casts.

Did not do

output_data[0] = 0 mutates the ORT tensor after Run. Works but surprising — float ort_score = entrance_found ? output_data[0] : 0.f; reads more naturally.

Did not do. I disagree that the latter reads more naturally than an explicit if statement

Nits:

SConscript indentation is inconsistent: surrounding lines use 2 spaces, new 'onnxruntime' uses 4 (mainlib) and 8 (plugins). Match surrounding style.

Did not do

_session_options and _allocator are only needed at construction. Could be locals. Minor.

Did not do

Dynamic-dim rewrite (if (dim == -1) dim = 1;) fixes batch=1. Batching N tracks per event would be a cheap win later — worth a TODO.

Added the TODO in ac29a48 but I'm not sure we can guarantee a set number of tracks in each event